Self-aware Pedestrians Modeling for Testing Autonomous Vehicles in

Simulation

Qazi Hamza Jan, Jan Markus Arnold Kleen and Karsten Berns

Robotics Research Lab, Technische Universität Kaiserslautern, 67663 Kaiserslautern, Germany

Keywords:

Pedestrian Simulation, Pedestrian Path, Autonomous-vehicles, Testing, Virtual Pedestrians.

Abstract:

With the rise of autonomous vehicles in the urban environment, the focus is also shifted towards autonomous

vehicles in pedestrian zones. Pedestrian safety becomes the primary concern in such zones. Autonomous

systems for these situations need thorough testing before its deployment in the real-world to ensure safety.

Therefore, developing testbeds that resemble the real-world for autonomous vehicles testing in pedestrian

zones are highly critical. The proposed work focuses on the modeling of pedestrian behaviors in a simulated

environment for realizing autonomous vehicles in the pedestrian zones. The virtual pedestrians are modeled

with the self-awareness to avoid static and dynamic obstacles when progressing towards its goal. The goal

is also to have a minimum number of parameters to generate various test scenarios with realistic behaving

pedestrians for the autonomous systems. The proposed system is evaluated using individual and group of

virtual pedestrians. It can be seen from the experiments that simulated pedestrians show trajectories which

resemble the trajectories of pedestrians in the real-world for that particular situation.

1 INTRODUCTION

The development of autonomous commute for the

pedestrian zone is on the rise. Many companies are

producing vehicles such as Navya

and Easymile

to provide solutions for transportation in pedestrian

zones. Pedestrian zones are public areas that in-

clude vast tourist attractions, airports, university cam-

puses, and so on. Because of environmental, so-

cial, and economic aspects, transportation authori-

ties are aiming at turning existing streets into pedes-

trian zones (Di

einait-Rauktien et al., 2018). How-

ever, walking such long distances becomes tiring, es-

pecially for the elderly and disabled people. There-

fore, there is a need to develop a safe and reliable

transportation system, such as driverless shuttles, that

can carry passengers without disrupting pedestrians

in such zones. The most critical aspect of an au-

tonomous vehicle in these areas is the safety of the

pedestrian. Safe navigation becomes critical because

of the random behavior of pedestrians. Since pedestri-

ans can include juveniles, seniors, and children, there-

fore, the autonomous transportation system should

make appropriate decisions according to the type of

pedestrian.

https://navya.tech/en/

https://easymile.com/

Proposed approaches for autonomous navigation

in the pedestrian zones should aim at guaranteeing

performance and, more importantly, safety. There-

fore, it should be tested thoroughly to drive safely

among people. Generally, researchers develop au-

tonomous transportation systems in a simulated en-

vironment to avoid any potential risks towards prop-

erty or the people. Despite the beneﬁts of develop-

ing vehicles in simulation, virtual pedestrians hardly

represent the actual behavior of real-life people. Hu-

mans express a variety of behaviors according to their

circumstances and personalities. Some of these be-

haviors can be instinctive, e.g., avoid collisions, and

some behaviors are goal-oriented, e.g., take the short-

est path. Additionally, some behaviors are based on

social norms, such as friends or families walking right

next to each other. Hence, these real-life behaviors

associated with different situations should also be in-

corporated into the simulation to develop a realistic

scenario in simulation.

The main contribution of this paper is to gener-

ate virtual pedestrians, called characters, that depict

real-life behaviors of pedestrians in pedestrian zones.

Every character which is spawned in the environment

is given knowledge about its surrounding. From this

knowledge, the characters are capable of reaching

their goal position by avoiding obstacles and vehicles

on their path. There are some user given parameters

Jan, Q., Kleen, J. and Berns, K.

Self-aware Pedestrians Modeling for Testing Autonomous Vehicles in Simulation.

DOI: 10.5220/0009377505770584

In Proceedings of the 6th International Conference on Vehicle Technology and Intelligent Transport Systems (VEHITS 2020), pages 577-584

ISBN: 978-989-758-419-0

577

to make the characters not to dodge the vehicles for

exceptional testing and validation of autonomous ve-

hicles themselves. The implementation is done in Un-

real Engine

In the next section, related work is presented,

which discusses similar pedestrian simulations al-

ready existing. Section 3 gives the overall architecture

of the system. In this section, the main components of

the proposed system are summarized. In section 4, the

system approach is discussed in detail. Finally, the re-

sults of the experiments are evaluated in section 5.

2 RELATED WORK

The study of autonomous vehicles in the pedestrian

zones is growing. Researchers are developing interac-

tion strategies to communicate with pedestrians. Jan

et al. (Jan et al., 2020) have developed interaction

modules to indicate pedestrians for avoiding colli-

sions with vehicles in advance. The authors have used

simulation to examine their results from the vehicle’s

perspective. Similarly, Rasouli and Tsotsos (Rasouli

and Tsotsos, 2019) have surveyed to identify factors

that affect the behavior of pedestrians and how to

solve interaction problems. They mention various de-

sign approaches to see the intentions of pedestrians.

For a thorough analysis of vehicle crowd interaction

in an urban environment, Yang et al. (Yang et al.,

2019) have proposed pedestrian trajectory datasets to

validate pedestrian motion models inﬂuenced by the

vehicles. The authors collect the trajectories of pedes-

trian and vehicle from a down-facing camera attached

to a drone. It particularly helps the community, which

is closely working in the interaction between pedes-

trians and vehicles.

There exists a well-known open-source simulation

environment, Carla (Dosovitskiy et al., 2017), which

is developed for autonomous driving research. Be-

sides vehicle driving and town modeling, Carla also

offers virtual pedestrians that navigate by location-

based cost. This cost encourages the pedestrian to

walk speciﬁcally on footpaths and main crossings.

However, Carla does not offer pedestrian behavior ex-

plicitly for vast pedestrian zones in which pedestrians

are not bounded by footpath like structures but can

walk freely along the stretch of the pedestrian zone

as individuals or groups. Another simulation sys-

tem speciﬁcally for representing the crowd is given

in (Kimura et al., 2019), in which the authors are us-

ing a multi-agent model for representing the crowd

in the SimTread. They have mainly analyzed evacu-

https://www.unrealengine.com/

ation cases and portray behaviors in bottleneck areas.

SimTread allows user-deﬁnable pedestrians and spa-

tial models. The overhead of using SimTread for any

environment is to have a spatial model of the scene.

Many different approaches have been imple-

mented for modeling character movements. Mehdi

Moussaïd et al. (Moussaïd et al., 2011) have ad-

dressed the issue of crowd disaster and modeled the

pedestrian ﬂow in simulation. Authors have used a

cognitive approach to adapt the walking speed and di-

rection of characters giving way to realistic modeling

of collective social behaviors. It uses vision to per-

ceive the environment, walking speed, and direction

as heuristics. A local method for collision avoidance

between characters has been done in (Karamouzas

et al., 2009), in which the collision is predicted,

and a smooth trajectory is calculated. The authors

have modeled the pedestrian as a disk on a 2D plane

with rotation. It mainly uses forces to calculate the

collision-free path until the goal position is reached.

In one of the recent work, data-driven approach has

been presented (Martin and Parisi, 2019) in which

experimental data has been collected from motion

capture sensors and fed into a generalized regression

neural networks. However, authors assume obstacle

avoidance between one pedestrian and a ﬁxed ob-

stacle. A classical approach for avoiding collision

with other bodies is presented in (van den Berg et al.,

2011). Their approach is based on velocity obsta-

cle (Fiorini and Shiller, 1998) for the collision-free

motion for different bodies. In most recent works, Yin

et al. (Yin et al., 2019) have used this approach for vir-

tual pedestrian simulation. However, the authors have

not differentiated pedestrian behavior for the case of

vehicles moving along their path. Inspired by pedes-

trian dynamics using social force models in (Helbing

and Molnar, 1995),(Helbing et al., 2000), (Alahi et al.,

2016), and (Cosgun et al., 2016), authors of (Chao

et al., 2019) also use scalable force-based framework

for simulating pedestrians bicycle and vehicles. Their

forces are dependent on the structured environment,

such as lanes and crossings. Our system offers a so-

lution for unstructured pedestrian zones where excep-

tional vehicles can also navigate.

3 SYSTEM ARCHITECTURE

The architecture of the proposed system is designed to

create several characters in the environment with their

behaviors. It also generates different scenarios with

a minimal number of parameters for testing different

aspects of autonomous vehicle systems. In pedes-

trian zones, plenty of behaviors are exhibited by the

VEHITS 2020 - 6th International Conference on Vehicle Technology and Intelligent Transport Systems

578

Figure 1: The overall architecture of the system. It consists

of three main modules in which every module deﬁnes its

function.

pedestrians, for instance, crossing in close vicinity of

the vehicle, walking in different group sizes in which

each group member has his distinct behavior, and so

on. Producing each of these behaviors for prolonged

test cases in the simulation is cumbersome. There-

fore, the system implemented in this paper also fo-

cuses on giving the user a simple API for generating

a variety of scenarios for testing interactivity between

pedestrians and autonomous vehicles. Some of these

parameters are included in the 4.6. The system is di-

vided into three main components; Pedestrian Control

Unit (PCU), Static Waypoint (SW), and Base Char-

acter (BC). Figure 1 shows the system architecture

with three components and their respective properties.

Each component is brieﬂy explained in the following

sub-sections.

3.1 Pedestrian Control Unit

PCU is designed for the systematic execution of the

characters and their behaviors. PCU manages all the

characters from spawning to despawning in the envi-

ronment. PCU directly provides access to different

parameters, which can be deﬁned by the user to gen-

erate different behaviors for characters. These param-

eters are brieﬂy given in 4.6. PCU entirely handles all

the work, from path making to managing characters

on these paths. The process is as follows:

• In the start of the simulation, PCU spawns the

characters in the environment. The spawning

could be as individual characters or group of char-

acters depending on the input from the user.

• Characters get a random path assigned to it and

spawned at the ﬁrst SW of the given path.

• The PCU also assigns different options given by

the user, e.g., number of characters, their group

size, and their speed.

• Assigning animations for each character, such as

talking on the phone, and texting.

3.2 Base Character

The entity to show the real-world pedestrians are the

BCs. BCs are the set of virtual pedestrians that move

along the path. The unreal engine includes highly-

featured Character class, which makes it walk, jump,

run, messaging, and talking on the phone. These par-

ticular set of animations plays an important role in

behavior and activity recognition of pedestrians in the

real world. A different attribute of characters such as

male/female, child/young, different complexions, and

size are included in order to fully exploit the percep-

tion system of autonomous vehicles. Each BC has its

individual movement. Obstacle avoidance is imple-

mented in BC. Therefore, each character uses its per-

ception to avoid an obstacle. Based on the available

path, it decides for its speed and direction.

3.3 Static Waypoint

The path must be deﬁned to have a meaningful strat-

egy for every BCs. It is done by placing SWs where

the characters are intended to walk. SW is the funda-

mental component for making paths. Each BC starts

from the initial SW and traverses all the SWs in a path

until it has reached the endpoint. At this point, ei-

ther it can select a new path or despawn based on the

information from the PCU. SW can be placed any-

where in the environment. However, for testing the

algorithms in an autonomous vehicle, SWs are placed

at the same locations where the vehicle has to navi-

gate. By slightly changing the location and number of

SW, different interactions between characters and ve-

hicles can be obtained. It is because characters need

to pass through every SW once, and if placed at the

same place as the vehicle, the characters can have a

waiting behavior until the SW is available to pass.

4 PROPOSED APPROACH

The main idea of this work is to provide real-life

pedestrian behaviors to characters in the simula-

tion environment for testing of autonomous systems.

These behaviors include moving towards a goal on a

suitable path and avoiding obstacles or vehicles on the

same path. For this reason, there is a need to deﬁne

heuristics for every character based on user parame-

ters. The approach builds on providing the surround-

ing knowledge to every character and deﬁning the set

of rules for speed and direction.

Self-aware Pedestrians Modeling for Testing Autonomous Vehicles in Simulation

579

Since the path for the characters is deﬁned at the

start of simulation by giving SWs, the focus here is

on collision avoidance, local pathﬁnding, and random

behaviors. Each character possesses its behavior and

visualization so that they depict their particular be-

havior. Additionally, the characters are also modeled

as individuals and in groups and how it reacts to au-

tonomous vehicles. In this paper, every walking char-

acter gets some features for a realistic reaction to au-

tonomous vehicles. These features are awareness, di-

rection heuristics, speed heuristics, SW identiﬁcation,

grouping, and can be modiﬁed by user-deﬁned param-

eters. All these features are explained in the subsec-

tion below.

4.1 Awareness

Rationally, it is not possible to avoid obstacles with-

out perceiving the environment. For realistic trajecto-

ries of characters, every character is equipped with, at

least, the basic ability to understand the environment.

It includes the ability to know if a moving vehicle

would be crossing their current direction. Addition-

ally, characters have the knowledge if the vehicle is

moving behind them and in the same direction. This

knowledge is given by using the "LineTrace" method

in Unreal Engine, which returns ﬁrst hit to an object

in the environment, as shown in ﬁgure 2. The width

of the person is also considered for obstacle avoid-

ance at every detection angle. Within already deﬁned

resolution for the ﬁeld of view, every possible walking

direction is checked. The additional direction towards

the target is considered in order to avoid the zig-zag

movement.

Figure 2: Visualization of a pedestrian c and its awareness

in simulation. α

directs towards the goal position. Green

area outlines the front of pedestrian whereas pink lines show

the rear of pedestrian, mainly, for capturing approaching ve-

hicle from behind.

4.2 Direction Heuristics

Every t seconds, a character has to re-evaluate their

direction. With the given information they can choose

between a certain, ﬁnite amount of directions. A dis-

tance value for each direction is calculated, based on

the given information and how direct the target is.

Then the direction with the smallest value is cho-

sen. This paper takes the minimization problem from

(Moussaïd et al., 2011) and modify and extends it:

d(α) = d

max

+ f (α)

− 2d

max

f (α)cos(α

− α) (1)

Where d is the (weighted) distance in direction α

between two end points which is shown in ﬁgure 3;

is the direction of the destination point; f (α) is

the distance to ﬁrst collision in direction α; and d

max

is the maximal viewing distance of a character.

First, d

max

is removed because it does not have an

effect on which direction has the smallest value. Ex-

periments with the heuristics determined that it works

well for small objects.

Next, 2d

max

f (α)cos(α

− α) causes the charac-

ters to only choose very direct directions, even if that

means walking against a wall. This is due to the fact

that every d(α), where (α

− α) > 90

◦

, is larger than

d(α

). But characters should be able to avoid not only

very small but also bigger obstacles. For this reason,

− α is valued less and 2d

max

f (α)cos(α

− α) is

replaced with 2d

max

f (α)cos(

−α

). In experiments

x = 2 seems to work quite well. This way, directions

with (α

− α) > 90

◦

have a chance to be picked. It is

noteworthy that avoidance of small objects does not

get worse because characters still choose the most di-

rect direction. Hence, the equation 1 is modiﬁed to

equation 2.

d(α) = f (α)

− 2d

max

f (α)cos(

− α

) (2)

Subsequently, characters should avoid walking in

front of or crossing paths with a vehicle, so the value

for these directions is increased by adding f (α)

equation 2. In other words, the distance to the colli-

sion is valued double. The reason to double is to avoid

the vehicle in advance, which ensures that this direc-

tion is not chosen. However, if all other directions

have a terrible distance to the collision, it still can be

chosen. If such a direction is chosen, characters may

want to have a different walking speed.

4.3 Speed Heuristics

After choosing a direction, characters have to choose

their walking speed. A constant minimum time to col-

lision t

coll

is set and used to calculate the character’s

speed. A preferred walking speed of an individual,

VEHITS 2020 - 6th International Conference on Vehicle Technology and Intelligent Transport Systems

580

Figure 3: The law of cosines applied to character move-

ment.

pre f

, is set by the user and used as an upper bound.

Equation 3 shows the speed of the character.

v = min(v

pre f

;

f (α)

coll

) (3)

Since t

coll

is constant and positive, the characters can

never walk against any object. The next contribu-

tion of this system is that the characters of this pa-

per should also try to avoid getting hit by vehicles. If

the chosen direction crosses the current direction of a

near moving vehicle, the character waits for the vehi-

cle to pass or stop, just like pedestrians crossing a road

in real life. For directions crossing a near moving ve-

hicle, v = 0, else v is chosen according to equation 3.

4.4 SW Identiﬁcation

A character constantly checks if it has reached its cur-

rent SW in the path. Reaching means the distance

to the SW is smaller than a distance ε (radius deﬁne

for every SW). When this happens, the character ei-

ther sets the next SW in its path as their current SW

or it has ﬁnished its current path. In the latter case,

the character either despawns from the environment

or queries for a suitable new path from the PCU de-

pending on given user settings in the PCU.

It is important to note that SWs should not

be placed in a way that they are blocked continu-

ously, but they can be blocked temporarily by an au-

tonomous vehicle. In this case, the simulation in-

creases the distance, ε, temporarily.

4.5 Grouping

One usual occurrence in real life is the pedestrians

walking in groups. There is no central decision-maker

for the characters in the group. Characters in a group

can still walk on their own, and they all have the same

Figure 4: The distance to group for group member 2 is ku −

vk.

SW at all times, as described previously. However,

they need to stay as a group. This requirement is

fulﬁlled by changing the speeds of characters in one

group.

First, the preferred walking speeds(ranging be-

tween 1.2 m/s to 3.6 m/s) for each character in the

group are equal, so without disruption, the group stays

together. It can still happen that members of the group

get so far away that it would not resemble a group

anymore. In this case, a single character needs to wait

for the rest of the group. Characters need to determine

when they should wait. Normally, people in a group

stay close together but are also comfortable with their

personal space. This concept is realized by using two

constants, a preferred and a maximum distance to the

group. Distance to a group, in this context, is mea-

sured by the difference in the personal distance to the

current SW and the maximum distance to the current

SW in the group, as shown in ﬁgure 4. Distance is

measured in this way to allow a group to split up in

order to walk around an object. This check always

guarantees that at least one member of the group is not

waiting for the group, hence, avoiding any deadlocks

caused by this functionality. If the distance between

the character and its group is greater than the thresh-

old distance, then the character waits for the group

to join by looking towards the rest of the group, sig-

naling they are waiting for them. As the distance to

the group becomes smaller or equal to the preferred

distance, they start walking at their regular speed. A

maximum distance is an upper bound to the preferred

distance to the group.

4.6 User Deﬁned Parameters

To generate each BC with different behavior is re-

ﬂected with the use of behavioral options that can ei-

ther apply to all or random characters. These options

are set by the PCU when spawning a BC according to

what parameters are set in the PCU.

As an example, one option that can be randomly

assigned to every character, which can be described as

irrational or “childlike”, is the ignoring of vehicles. It

can be split into two different options. First, ignor-

ing if a direction crosses the path of a vehicle when

choosing it and second, the option to not wait for ve-

hicles. The ﬁrst only affects the direction, the second

Self-aware Pedestrians Modeling for Testing Autonomous Vehicles in Simulation

581

(a) Scenario 1 (b) Scenario 2

Figure 5: Two 2D-time plots showing the trajectory of characters. Plot(a) has two characters walking in the direction of the

vehicle. The start and end of vehicle path is shown with green dot and red cross respectively. Red asterisk shows the SW. Gray

area shows the pedestrian path. Plot(b) depicts trajectories of ﬁve characters walking in a group avoiding a static obstacle(in

red).

only speed. It is implemented to reﬂect reality, where

not all characters behave rationally.

Moreover, there are a series of more technical op-

tions. These include the already mentioned maximum

group size, the delay between spawns, the distance

epsilon for reaching an SW, and if characters should

despawn on ﬁnishing a path. The rest of the param-

eters include character speed, which is a range from

walking to running, animation chance, for having ran-

dom animation upon reaching an SW.

5 EXPERIMENTS AND

EVALUATION

In real-world pedestrian zones, there exists an enor-

mous number of situations where pedestrians behave

differently. This variation in behavior depends on the

number of pedestrians, their age group, their motives,

awareness level, and many more.

To exploit every behavior of pedestrians existing

in the real-world is impractical in the length of the

paper. For this reason, three particular scenarios are

included, which subsume additional behaviors occur-

ring in the pedestrian zone. The scenario consists of

a campus-like environment modeled in the simulation

where there are no speciﬁc road markings, and pedes-

trians are expected to walk anywhere on the pathway.

These scenarios are plotted using 2D time graphs of

ground truth values of characters and also, later, by

showing simulation images to prove the discernabil-

ity of the implemented system.

In the ﬁrst scenario, a vehicle crossing two char-

acters is presented. The plot for this experiment is

shown in ﬁgure 5a. The two characters are given

two SWs (shown with a red asterisk) for the start and

goal position. In an ordinary situation where there is

no obstacle, the characters walk straight towards the

goal position. But as a vehicle approaches towards

them, the characters start making way for the vehicle

to pass. This behavior can be seen by inspecting the

graph from "-23" Y-position onwards. Both charac-

ters change the course of their way. It can be seen

from the vehicle trajectory with a start as a green dot

and end as a red cross that the vehicle is slightly made

to steer towards the characters. By doing so, the char-

acters have reached the end of the width of the pedes-

trian path sketched as a gray parallelogram. In this

experiment, both the vehicle and the characters are

moving with walking speed.

A group of ﬁve characters is designated by deﬁn-

ing the parameters in the PCU for the second sce-

nario to evaluate the grouping behavior of the pro-

posed system. To justify their resemblance in their

trajectory, a static obstacle is placed between their

goal and end positions, as can be seen in ﬁgure 5b.

The plot shows the trajectory of the group avoiding

the obstacle. From the trajectories of every charac-

ter, it can be seen how closely they remain throughout

their path. This behavior deﬁnes the characteristics of

a group. At some point in the plot, the uniformity of

the group is lacking. However, considering the same

in a real-world scenario, uniformity does not always

exist. People incline in different directions within the

group. For example, if someone stops, the person be-

hind would try to cross him/her.

The ﬁnal scenario is to illustrate the behavior of

the group for a crossing vehicle. It is shown by us-

ing images directly from our model in the simulation.

The environment is shown in ﬁgure 6. The trajec-

tory traversed by each character in the group clariﬁes

VEHITS 2020 - 6th International Conference on Vehicle Technology and Intelligent Transport Systems

582

(a) Trajectory 1 (b) Trajectory 2

Figure 6: Group of six characters walking towards an approaching vehicle in a narrow pedestrian zone.

that giving way to an oncoming vehicle is more im-

portant than staying in a group. The vehicle is made

intentionally to move among the group to observe the

group’s reaction. The characters are closely grouped

when the vehicle is far away from them, as can be

seen in ﬁgure 6a. As the vehicle approaches towards

the group of characters, ﬁgure 6b shows the characters

start making space for the vehicle to pass. Since there

is no other way for the vehicle to evade the whole

group of characters from both the sideways, the group

divides in a reasonable way to avoid confusion, as can

be viewed in ﬁgure 6c. Finally, ﬁgure 6d spots that the

group integrates back once there is no other obstacle

on the way.

The system validation can be done by comparing

the trajectories from the plots in this section to real-

world pedestrians. In a real-world situation, similar

characteristics of the splitting of groups can be seen.

However, the trajectories of pedestrians may differ for

every individual. Our experiments show trajectories

similar to shown in(Appert-Rolland et al., 2018), who

have reported the study of pedestrian trajectories as

individuals and groups. They have shown pattern for-

mation of a pedestrian in a different medium of crowd

density.

6 CONCLUSION

This paper proposes and implements virtual pedestri-

ans in simulation, which carry real-world pedestrian

like behavior. From the experiments, it can be in-

spected that characters for different instances show

reasonable behaviors, which is also expected in the

real-world. The plots in ﬁgure 5, demonstrate that

the characters distinguish between static and dynamic

obstacle. The characters in plot 5a constantly change

in their trajectory as the vehicle approaches them, but

the characters in plot 5b pass around the obstacle to

follow the shortest distance towards the goal. These

behaviors are normally expected in humans when they

want to cross a vehicle or wall in their path.

It is understood that humans use different gestures

for communication to ﬁnd a consensus for their pref-

erence while crossing other people or vehicles. The

characters in the proposed system only use visual aid

to avoid obstacles. They can be enhanced by adding

hearing aid and signaling gestures as well in future

work. The combined effect of these aids can help in

better realizing the interaction and trajectory between

the characters and the vehicle.

Self-aware Pedestrians Modeling for Testing Autonomous Vehicles in Simulation

583

REFERENCES

Alahi, A., Goel, K., Ramanathan, V., Robicquet, A., Fei-

Fei, L., and Savarese, S. (2016). Social lstm: Human

trajectory prediction in crowded spaces. In Proceed-

ings of the IEEE conference on computer vision and

pattern recognition, pages 961–971.

Appert-Rolland, C., Pettré, J., Olivier, A.-H., Warren, W.,

Duigou-Majumdar, A., Pinsard, É., and Nicolas, A.

(2018). Experimental study of collective pedestrian

dynamics. arXiv preprint arXiv:1809.06817.

Chao, Q., Jin, X., Huang, H.-W., Foong, S., Yu, L.-F., and

Yeung, S.-K. (2019). Force-based heterogeneous traf-

ﬁc simulation for autonomous vehicle testing. In 2019

International Conference on Robotics and Automation

(ICRA), pages 8298–8304. IEEE.

Cosgun, A., Sisbot, E. A., and Christensen, H. I. (2016).

Anticipatory robot path planning in human environ-

ments. In 2016 25th IEEE International Sympo-

sium on Robot and Human Interactive Communica-

tion (RO-MAN), pages 562–569. IEEE.

einait-Rauktien, R., Val

eiukien, J., Parsova, V., and

Maliene, V. (2018). The importance of environmen-

tal criteria for kaunas city pedestrian zones. Opportu-

nities and Constraints of Land Management in Local

and Regional Development: Integrated Knowledge,

Factors and Trade-offs, page 133.

Dosovitskiy, A., Ros, G., Codevilla, F., Lopez, A., and

Koltun, V. (2017). CARLA: An open urban driving

simulator. In Proceedings of the 1st Annual Confer-

ence on Robot Learning, pages 1–16.

Fiorini, P. and Shiller, Z. (1998). Motion planning in dy-

namic environments using velocity obstacles. The In-

ternational Journal of Robotics Research, 17(7):760–

772.

Helbing, D., Farkas, I., and Vicsek, T. (2000). Simu-

lating dynamical features of escape panic. Nature,

407(6803):487.

Helbing, D. and Molnar, P. (1995). Social force model for

pedestrian dynamics. Physical review E, 51(5):4282.

Jan, Q. H., Klein, S., and Berns, K. (2020). Safe and efﬁ-

cient navigation of an autonomous shuttle in a pedes-

trian zone. In Berns, K. and Görges, D., editors, Ad-

vances in Service and Industrial Robotics, pages 267–

274, Cham. Springer International Publishing.

Karamouzas, I., Heil, P., van Beek, P., and Overmars, M. H.

(2009). A predictive collision avoidance model for

pedestrian simulation. In Egges, A., Geraerts, R., and

Overmars, M., editors, Motion in Games, pages 41–

52, Berlin, Heidelberg. Springer Berlin Heidelberg.

Kimura, T., Sano, T., Hayashida, K., Takeichi, N.,

Minegishi, Y., Yoshida, Y., and Watanabe, H. (2019).

Representing crowds using a multi-agent model–

development of the simtread pedestrian simulation

system. Japan Architectural Review, 2(1):101–110.

Martin, R. F. and Parisi, D. R. (2019). Data-driven simula-

tion of pedestrian collision avoidance with a nonpara-

metric neural network. Neurocomputing.

Moussaïd, M., Helbing, D., and Theraulaz, G. (2011). How

simple rules determine pedestrian behavior and crowd

disasters. Proceedings of the National Academy of

Sciences, 108(17):6884–6888.

Rasouli, A. and Tsotsos, J. K. (2019). Autonomous vehi-

cles that interact with pedestrians: A survey of theory

and practice. IEEE Transactions on Intelligent Trans-

portation Systems.

van den Berg, J., Guy, S. J., Lin, M., and Manocha, D.

(2011). Reciprocal n-body collision avoidance. In

Pradalier, C., Siegwart, R., and Hirzinger, G., editors,

Robotics Research, pages 3–19, Berlin, Heidelberg.

Springer Berlin Heidelberg.

Yang, D., Li, L., Redmill, K., and Özgüner, Ü. (2019).

Top-view trajectories: A pedestrian dataset of vehicle-

crowd interaction from controlled experiments and

crowded campus. arXiv preprint arXiv:1902.00487.

Yin, Z., Liu, J., and Wang, L. (2019). Less-effort colli-

sion avoidance in virtual pedestrian simulation. In

Proceedings of the 2019 International Conference on

Artiﬁcial Intelligence and Computer Science, AICS

2019, pages 488–493, New York, NY, USA. ACM.

VEHITS 2020 - 6th International Conference on Vehicle Technology and Intelligent Transport Systems

584